Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Better, Faster, Stronger: Programmatic Skill Learning Best Reduces Agent Cost

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper argues that representing an agent's learned skills as executable code, rather than natural-language notes, is the skill-learning strategy that best reduces inference cost, and demonstrates the point with SpeedRunner across three…

desk verdict The cost-reduction result is credible and well-measured, but the 'programs beat prose' headline outruns the baselines. read the letter →

arxiv 2608.11338 v1 pith:RIJY4VOH submitted 2026-08-11 cs.CL cs.LG

Zixi Huang , Xiheng Wang , Andrew Wang , William Jurayj , Bernal Jiménez Gutiérrez , Daniel Khashabi , Nicholas Andrews This is my paper · ORCID
classification cs.CLcs.LG
keywords programmaticskillsskilllearningcostreductionLLMagentstrajectoryanalysiswake-sleeplibrarycodeinterpreterembodied
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that, among ways of teaching an LLM agent reusable skills, writing those skills as executable functions achieves the largest reduction in inference cost, while natural-language skill notes keep charging for the same reasoning every time they are used. To test this, the authors build SpeedRunner, an agent that alternates between acting with a skill library and sleeping: in the sleep phase, a coding agent with a code interpreter inspects raw stored trajectories, diagnoses repeated failures, and edits library functions. SpeedRunner is evaluated on three text-based embodied environments and is the only method whose per-episode cost falls as training proceeds; on BabyAI it drops to about an eighth of the ReAct no-learning baseline while performance climbs to near-perfect. The practical stakes are that deployed agents could adapt online, without replay buffers, environment rewinding, or held-out validation, and get cheaper as they gain experience.

What carries the argument

The central object is a wake-sleep skill-library loop. A stochastic policy (the actor) is equipped with a library of executable skills, each a function plus documentation over the environment's atomic actions; during the wake phase it rolls out trajectories, and during the sleep phase an inducer—a coding agent with access to a code interpreter—adds, edits, or deletes library entries. Trajectories are stored in full with call stacks and tags identifying which library version generated them, so the inducer can query specific slices of history instead of reading everything into context. Public/private access modifiers keep helper functions callable by other skills while hiding them from the actor, bounding the library size the actor must attend to. This machinery is what converts raw, noisy traces into reusable, compositional routines rather than accumulated special cases.

What would settle it

Run SpeedRunner on BabyAI with the inducer's code interpreter disabled while keeping the actor, budget, and library mechanics identical; if per-episode cost still falls to around an eighth of the ReAct baseline and reversed-order mission failures do not reappear, then programmatic trajectory analysis is not what drives the compression. A cheaper probe is to inspect a no-interpreter library for the priority-ordered mission parser the paper identifies as the decisive fix.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the main measurable payoff of programmatic skill learning is cost: representing a recurring routine as code lets the agent delegate reasoning to deterministic execution, so the routine is reasoned about once and then invoked cheaply, whereas prose skills must be reread and re-followed on every use. SpeedRunner shows that such skills can be induced purely online: its inducer treats trajectory history as a queryable dataset, using code to count failing skill calls, test candidate parsers against recorded observations, and refactor the library in place. The learned libraries are compact and hierarchical, with functions calling other functions, rather than append-only piles of scenario-specific snippets. On BabyAI the actor completes episodes with one or two high-level calls, and across Crafter, ScienceWorld, and BabyAI SpeedRunner is the only method whose cost decreases over training.

Load-bearing premise

The load-bearing premise is that raw past trajectories contain enough signal for a coding-agent inducer to learn reliable, reusable routines by inspection alone, with no replay or validation; if noisy long-horizon traces do not carry that signal, the cost advantage collapses.

Editorial extensions

If this is right

  • An agent can gain environment-specific skills online, on ephemeral tasks, without replay, validation, or the ability to rewind the environment.
  • Per-task inference cost should fall as the library matures, because recurring behavior is compressed into deterministic functions instead of re-derived each episode.
  • Programmatic skill learning should survive environmental randomness and task-distribution shifts, as long as the inducer can edit existing skills in place rather than only append new ones.
  • The amount of performance gain will vary with how much reusable procedure the environment exposes: high on BabyAI and ScienceWorld, more modest on stochastic Crafter, but the cost compression persists.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A conservative reading of the cross-model results is that code delegation is a cost-performance dial: with stronger base models, aggressive compression can make the policy too rigid, so future systems might expose a tuning knob controlling how much decision-making is handed to code.
  • Since the mechanism is generic trajectory inspection, the same design could be tried in non-embodied agent settings (tool use, web navigation, customer service) where raw interaction logs already exist; the paper only evaluates embodied text worlds.
  • A testable extension is to add a cheap static regression gate—for example, prevent a library edit if call-graph depth or recent failure counts move the wrong way—since the paper reports instability across all methods and leaves stabilizing online induction as an open problem.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper investigates cost-effective online skill learning for LLM agents and argues that skills represented as executable code yield the largest reduction in inference cost among skill-learning strategies, since deterministic program execution replaces repeated LLM reasoning. It introduces SpeedRunner, a wake-sleep algorithm in which an actor, equipped with a library of callable skills, generates trajectories, and an inducer—a coding agent with a Python interpreter—analyzes the trajectory history programmatically to add, edit, and delete functions. Experiments on ScienceWorld, Crafter, and BabyAI compare SpeedRunner against ReAct (no learning), OPO (natural-language prompt optimization over concatenated trajectories), ASI (append-only code skills), and Voyager (append-only code library, on Crafter), with GPT-5.4-mini as the default backbone and Gemini-3-Flash and Qwen-3.5-27B as cross-model checks. The central empirical claim is that SpeedRunner consistently reaches the performance–cost frontier, with cost decreasing over training, including a roughly eightfold cost reduction on BabyAI relative to ReAct. Additional analyses cover ablations of the inducer components, call-graph structure of learned libraries, robustness under Crafter zombie-frequency variation, and adaptation to a ScienceWorld distribution shift.

Significance. The paper's core finding—that an editing, compressing library of executable skills reduces per-episode cost over training, and does so across three benchmarks and three backbones—is credible and carefully measured. Strengths include matched held-out test sets shared across methods, cost accounting that amortizes sleep-phase inducer costs, a post-hoc Voyager truncation experiment (Appendix K) that tests and rejects a context-bloat explanation for Voyager's regression, and unusually candid reporting in Appendix G.2 of settings where OPO beats SpeedRunner on final performance. If the comparative claim is the contribution, it is currently only partially established: the experiments support SpeedRunner over the tested baselines, but they do not yet support the paper's broader claim of superiority of program representation over all skill-learning methods. The cross-model checks partially mitigate the acknowledged limitation of not testing flagship models, but they also expose a performance-claim inconsistency discussed below.

major comments (3)
  1. [§4.1, §5 (Figure 3), §7] The headline claim—'among all the different skill learning methods, those that view skills as programs can achieve the best cost reduction' (Abstract)—is a comparison across skill-learning methods, but the only natural-language baseline in the main experiments, OPO, is not a distilling skill learner: per §4.1 it 'reviews a concatenation of recent trajectories and updates the actor prompt,' so its actor context grows with history and its Crafter cost increase largely reflects context bloat. The code baselines (ASI, Voyager) are append-only, so the code-vs-prose contrast is confounded with library curation. Meanwhile §7 cites natural-language skill-learning methods (Trace2Skill, SkillClaw, Agent Workflow Memory, SkillX, ExpeL) that summarize and curate trajectories into bounded-context documents and could plausibly retain the cost advantage of a curated representation. As it stands, the experiments establish that SpeedRunner's editing-and-compression design beats raw-history prompting and append-only libraries, but they do not isolate code representation per se as the cause of the cost frontier. The authors should either add a distilled natural-language skill baseline (e.g., Trace2Skill-style or Agent Workflow Memory-style prompt library with bounded context) or explicitly narrow the claim to programmatic skill learning with library curation.
  2. [§5 (Performance) vs §G.2 (Figures 9–10)] The results section overstates performance relative to the paper's own cross-model appendix. §5 states that 'SpeedRunner significantly outperforms all baselines across benchmarks in terms of performance and cost except for OPO in ScienceWorld on performance, per two-sided paired t-tests,' and that 'SpeedRunner achieves the strongest final performance on all three benchmarks.' Appendix G.2 reports the opposite for the other backbones: with Gemini-3-Flash, SpeedRunner achieves a significantly worse final success rate than OPO in all three benchmarks, and with Qwen-3.5-27B in ScienceWorld, despite much lower token usage. These statements must be qualified as model-specific (GPT-5.4-mini) or reconciled with the cross-model figures; as written, the main text contradicts the appendix, and the claim 'significantly outperforms' is misleading.
  3. [§4.3, §5] The statistical support for 'significantly outperforms' is not reported. All comparisons rest on three seeds (§4.3), and seed-level metrics are evaluated on a fixed 30-episode test set; paired t-tests over three paired observations have very low power, yet no p-values, effect sizes, or test details are given anywhere in the paper. Given that the headline contribution is the cost-reduction trend—which is clear and consistent across benchmarks and models—I recommend either reporting the test statistics or softening the significance language to 'consistently on the performance–cost frontier.'
minor comments (6)
  1. [Appendix E] The phrase 'purly-online setting' contains a typo and should read 'purely-online setting.'
  2. [Figure 2] The figure caption contains the placeholder text 'This is the methods figure,' which appears to be an editorial remnant and should be removed.
  3. [References / §7] The author name 'Huan ang Gao' in the reference list and the in-text citation 'ang Gao et al., 2026' appear corrupted; the intended surname is likely 'Ang Gao,' and the bibliographic entry should be corrected.
  4. [§4.3 / Figure 3] §4.3 states that efficiency is measured as 'output tokens per episode,' but Figure 3 shows dollar-denominated cost axes, and Appendix A.3 states that dollar-cost accounting is used in Figure 3; the main text should state the reported metric consistently.
  5. [Figure 1] The caption uses 'GPT 5.4-mini' while the rest of the paper uses 'GPT-5.4-mini'; the naming should be unified.
  6. [§6.2.2] The text says 'Two patterns recur across all four environments,' but the paper evaluates three benchmarks; if the two ScienceWorld tasks are counted as separate settings, this counting should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SpeedRunner's cost and performance claims are measured outcomes, not constructed from their inputs.

full rationale

The paper's central claim—that programmatic skill learning gives the best cost reduction—is an empirical comparative result supported by directly measured rollout costs and held-out success rates, not by a derivation in which the conclusion is built into the definition of the method or baseline. Skill libraries are induced from training trajectories and evaluated on disjoint test episodes with fixed checkpoints; cost is computed from token prices applied to actual LLM calls, including amortized inducer overhead. No fitted parameter is renamed as a prediction, and no uniqueness theorem or prior result by the same authors is invoked to force the design. The only overlapping self-citation (Wang et al., 2025a, cited for the claim that unbounded libraries degrade actor performance) is peripheral: it motivates the public/private split, and the paper independently observes the same context-overflow effect in ASI and Voyager. Concerns that OPO is not a distilled prose baseline are a comparison-fairness issue, not a circularity of the derivation chain. Under the quoted-evidence standard, no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and rests on a handful of hand-set experimental hyperparameters (sleep interval, budget, episode count, test size) and domain assumptions about price stability and representativeness. No new physical or mathematical entities are introduced.

free parameters (4)
  • sleep_cycle_interval = 10 rollouts
    Inducer invoked every 10 rollouts; chosen by hand, not optimized. Affects how frequently the library is updated and contributes to the measured cost.
  • training_episodes = 200
    All runs use 200 online rollouts; the paper does not study sensitivity to this budget.
  • actor_request_limit = 100 (ScienceWorld/BabyAI), 2000 (Crafter)
    Caps LLM calls per episode to bound cost; chosen because additional calls did not improve success, which affects both performance and cost.
  • test_set_size = 30
    Held-out episodes per benchmark; small size inflates variance of reported means.
assumptions (4)
  • domain assumption The three benchmarks (ScienceWorld, Crafter, BabyAI) are representative of embodied LLM-agent domains.
    Supports the paper's generalization claim; neither the environments nor the task families are chosen by an objective criterion.
  • domain assumption API prices published on May 6, 2026 are accurate and stable proxies for inference cost.
    All cost comparisons use these prices; if prices change, the magnitude of the cost advantage changes.
  • domain assumption The environment is a POMDP with tasks sampled from distribution T, and the agent has access to the listed primitive actions.
    Standard setup inherited from prior work; no proof that this models all deployment settings.
  • domain assumption The inducer's code interpreter faithfully reflects the environment's dynamics when it analyzes trajectories.
    The method relies on the inducer querying trajectories and testing hypotheses; if the code interpreter's view differs from the real environment, edits could be worse.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Better, Faster, Stronger: Programmatic Skill Learning Best Reduces Agent Cost." pith.science (2026). https://pith.science/paper/RIJY4VOH

@misc{pith2026260811338,
  author       = {Pith},
  title        = {Pith review of: Better, Faster, Stronger: Programmatic Skill Learning Best Reduces Agent Cost},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RIJY4VOH}},
  note         = {Machine review of arXiv:2608.11338}
}
read the original abstract

Recently, the practice of augmenting LLM agent capability with skills has gained prevalence. We explore the cost effective adaptation of agents to novel domains by means of learning skills. Existing works focus on performance gain over cost effectiveness. As a result, little is known about what skill learning strategies save cost. We argue that among all the different skill learning methods, those that view skills as programs can achieve the best cost reduction. By executing sequences of actions deterministically, a program-augmented agent can reliably and cheaply achieve goals that would otherwise require trial and error and risk degenerate behavior over long horizons. An agent can learn at inference time by incrementally discovering these programs and equipping them for future tasks. We hypothesize that past trajectories contain enough signal to guide skill learning, even without replay or validation, provided the agent can learn to analyze them. To test our claims, we propose SpeedRunner, a coding agent that analyzes trajectories and refactors skills for better performance on future tasks. Across three different embodied environments, we show that SpeedRunner consistently achieves the frontier in learning and cost reduction while remaining robust against distribution shifts and environmental randomness.

Figures

Figures reproduced from arXiv: 2608.11338 by the authors.

Figure 1
Figure 1. Performance vs average cost per trace for different approaches with GPT 5.4-mini on 3 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of SpeedRunner. A stochastic policy πL (the actor) is equipped with a skill library L, while a proposal function g (the inducer) updates the library using trajectory history H. During the wake phase, the actor generates a batch of trajectories. During the sleep phase, the inducer examines the trajectories and edits the library. In this example, the inducer notices repeated “not found” responses from travel_… view at source ↗
Figure 3
Figure 3. Performance and output-token cost across Crafter, ScienceWorld, and BabyAI. Our method [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Ablating SpeedRunner components on the BabyAI environment. Individual components have only small effects on SpeedRunner’s performance, however, removing the code interpreter degrades efficiency considerably. 6 Discussion 6.1 Ablation Study We ablate three components of…
Figure 5
Figure 5. Figure 5: Crafter final performance vs %Combat, the fraction of induced functions whose body references a creature/combat keyword under different zombie settings (GPT-5.4-mini, 200 training rollouts). 0 100 200 300 400 # of Training Traces 20% 40% 60% 80% Performance ( ) 3 4 0 1…
Figure 6
Figure 6. Figure 6: ScienceWorld distribution shift: performance (left) and cost (right) when switching from [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Extra SpeedRunner ablations on the Crafter and ScienceWorld environments. As in BabyAI, removing access to the code interpreter has only small effects on performance but contributes substantially to improvements in efficiency. 0 50 100 150 200 # of Training Traces 10% …
Figure 8
Figure 8. Figure 8: Main results using GPT-5.4-mini with tokens instead of cost for easy comparison across [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Gemini-3-Flash performance and output-token cost across Crafter, ScienceWorld, and [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Qwen-3.5-27B performance and output-token cost across Crafter, ScienceWorld, and [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Performance vs output tokens per trace for all approaches and benchmarks for GPT [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Effect of ASI’s replay-based skill verification on ScienceWorld task 3. Restoring replay [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 13 canonical work pages

  1. [1]

    Wang, Zora Zhiruo and Mao, Jiayuan and Fried, Daniel and Neubig, Graham , month = jun, year =. Agent

    work page

  2. [2]

    doi:10.48550/arXiv.2604.08377 , abstract =

    Ma, Ziyu and Yang, Shidong and Ji, Yuxiang and Wang, Xucong and Wang, Yong and Hu, Yiming and Huang, Tongwen and Chu, Xiangxiang , month = apr, year =. doi:10.48550/arXiv.2604.08377 , abstract =

    arXiv doi:10.48550/arxiv.2604.08377

  3. [3]

    Transactions on Machine Learning Research , author =

    Voyager:. Transactions on Machine Learning Research , author =. 2023 , keywords =

    work page 2023

  4. [4]

    , month = dec, year =

    Ellis, Kevin and Morales, Lucas and Sablé-Meyer, Mathias and Solar-Lezama, Armando and Tenenbaum, Joshua B. , month = dec, year =. Library learning for neurally-guided. Proceedings of the 32nd

    work page

  5. [5]

    , month = jun, year =

    Ellis, Kevin and Wong, Catherine and Nye, Maxwell and Sablé-Meyer, Mathias and Morales, Lucas and Hewitt, Luke and Cary, Luc and Solar-Lezama, Armando and Tenenbaum, Joshua B. , month = jun, year =. Proceedings of the 42nd. doi:10.1145/3453483.3454080 , abstract =

    arXiv

  6. [6]

    doi:10.48550/arXiv.2401.16467 , abstract =

    Stengel-Eskin, Elias and Prasad, Archiki and Bansal, Mohit , month = jun, year =. doi:10.48550/arXiv.2401.16467 , abstract =

    arXiv doi:10.48550/arxiv.2401.16467

  7. [7]

    2025 , eprint=

    Inducing Programmatic Skills for Agentic Tasks , author=. 2025 , eprint=

    2025

  8. [8]

    2025 , eprint=

    SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills , author=. 2025 , eprint=

    2025

Show all 40 references
  1. [9]

    2026 , eprint=

    GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning , author=. 2026 , eprint=

    2026

  2. [10]

    Gradient Descent

    Automatic Prompt Optimization with "Gradient Descent" and Beam Search , author=. 2023 , eprint=

    2023

  3. [11]

    Second Conference on Language Modeling , year=

    Hell or High Water: Evaluating Agentic Recovery from External Failures , author=. Second Conference on Language Modeling , year=

  4. [12]

    S cience W orld: Is your Agent Smarter than a 5th Grader?

    Wang, Ruoyao and Jansen, Peter and C \^o t \'e , Marc-Alexandre and Ammanabrolu, Prithviraj. S cience W orld: Is your Agent Smarter than a 5th Grader?. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.775

    2022 doi

  5. [13]

    International Conference on Learning Representations , year=

    Benchmarking the Spectrum of Agent Capabilities , author=. International Conference on Learning Representations , year=

  6. [14]

    Maxime Chevalier-Boisvert and Dzmitry Bahdanau and Salem Lahlou and Lucas Willems and Chitwan Saharia and Thien Huu Nguyen and Yoshua Bengio , booktitle=. Baby. 2019 , url=

    2019

  7. [15]

    The Eleventh International Conference on Learning Representations , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. The Eleventh International Conference on Learning Representations , year=

  8. [16]

    2024 , url=

    Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan , booktitle=. 2024 , url=

    2024

  9. [17]

    The Fourteenth International Conference on Learning Representations , year=

    Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces , author=. The Fourteenth International Conference on Learning Representations , year=

  10. [18]

    Nature , volume=

    Optimizing generative AI by backpropagating language model feedback , author=. Nature , volume=

  11. [19]

    Grounding large language models in interactive environments with online reinforcement learning , year =

    Carta, Thomas and Romac, Cl\'. Grounding large language models in interactive environments with online reinforcement learning , year =. Proceedings of the 40th International Conference on Machine Learning , articleno =

  12. [20]

    The Thirteenth International Conference on Learning Representations , year=

    Davide Paglieri and Bart. The Thirteenth International Conference on Learning Representations , year=

  13. [21]

    2026 , eprint=

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills , author=. 2026 , eprint=

    2026

  14. [22]

    2026 , eprint=

    AutoRefine: From Trajectories to Reusable Expertise for Continual LLM Agent Refinement , author=. 2026 , eprint=

    2026

  15. [23]

    2026 , eprint=

    SkillX: Automatically Constructing Skill Knowledge Bases for Agents , author=. 2026 , eprint=

    2026

  16. [24]

    2026 , eprint=

    EvoSkill: Automated Skill Discovery for Multi-Agent Systems , author=. 2026 , eprint=

    2026

  17. [25]

    2026 , eprint=

    SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support , author=. 2026 , eprint=. doi:https://doi.org/10.1145/3805712.3808466 , url=

    2026

  18. [26]

    2026 , eprint=

    SkillLearnBench: Benchmarking Continual Learning Methods for Agent Skill Generation on Real-World Tasks , author=. 2026 , eprint=

    2026

  19. [27]

    2025 , eprint=

    RLAD: Training LLMs to Discover Abstractions for Solving Reasoning Problems , author=. 2025 , eprint=

    2025

  20. [28]

    Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric

    Qirui Mi and Zhijian Ma and Mengyue Yang and Haoxuan Li and Yisen Wang and Haifeng Zhang and Jun Wang , booktitle=. Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric. 2026 , url=

    2026

  21. [29]

    2026 , eprint=

    Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents , author=. 2026 , eprint=

    2026

  22. [30]

    2026 , eprint=

    AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution , author=. 2026 , eprint=

    2026

  23. [31]

    and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and Hassabis, Demis and Clopath, Claudia and Kumaran, Dharshan and Hadsell, Raia , year=

    Kirkpatrick, James and Pascanu, Razvan and Rabinowitz, Neil and Veness, Joel and Desjardins, Guillaume and Rusu, Andrei A. and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and Hassabis, Demis and Clopath, Claudia and Kumaran, Dharshan and Ha...

    doi

  24. [32]

    2017 , eprint=

    Learning without Forgetting , author=. 2017 , eprint=

    2017

  25. [33]

    2024 , eprint=

    A Comprehensive Survey of Continual Learning: Theory, Method and Application , author=. 2024 , eprint=

    2024

  26. [34]

    2024 , eprint=

    Continual Learning of Large Language Models: A Comprehensive Survey , author=. 2024 , eprint=

    2024

  27. [35]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

    2023

  28. [36]

    2024 , eprint=

    ExpeL: LLM Agents Are Experiential Learners , author=. 2024 , eprint=

    2024

  29. [37]

    2026 , eprint=

    A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence , author=. 2026 , eprint=

    2026

  30. [38]

    2025 , eprint=

    A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems , author=. 2025 , eprint=

    2025

  31. [39]

    2024 , eprint=

    Trial and Error: Exploration-Based Trajectory Optimization for LLM Agents , author=. 2024 , eprint=

    2024

  32. [40]

    2025 , eprint=

    RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning , author=. 2025 , eprint=

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.